Design layered memory and persistent prefetch cache

[RenanGBarreto]

Why

Critical metadata should remain in memory when practical, while larger file-region data may eventually need persistent disk storage. Current implementation has no RemotePFS-owned disk cache.

Plan

• Define cache-layer requirements without adding an unimplemented cache protocol.
• Keep current executor reading through NBD/mapper and nbdkit cache.
• Measure metadata memory use and nbdkit temporary-file behavior first.
• Design memory-to-disk promotion only after a real RemotePFS-owned backend exists.
• Define ownership, generation identity, eviction, restart recovery, and invalidation semantics before implementation.

Acceptance criteria

• No parallel cache coherence or invalidation path is introduced prematurely.
• Design documents memory, future disk, and NAS fallback layers.
• Cache ownership and generation boundaries are explicit.
• Prototype includes measurements and restart/reload behavior before production use.